Standard 12-lead ECGs provide a multi-perspective view of cardiac electrophysiology by capturing 12 signals that encode critical information on the propagation of cardiac electrical activity. However, training models on multilabel ECG datasets presents a major challenge: the inherent imbalance between common and rare diseases leads to suboptimal performance, particularly on rare pathologies. One approach to mitigating class imbalance is to over-sample instances associated with rare diseases. While balanced-sampling is straightforward in multiclass settings, sampling becomes more complex in multilabel classification due to the presence of multiple labels per data point.
To address this, this work proposes a learnable sampling framework for multilabel classification. Unlike heuristic samplers in the literature that rely on discrete instance duplication, this approach formulates sampling as a continuous constrained optimization problem, with the solution being a probability distribution over instances that guides the training sampling process. The framework dynamically adjusts sampling probabilities to ensure balanced exposure across all disease classes. The method integrates class-specific prioritization directly into the optimization objective, giving clinicians the flexibility to focus training on critical (e.g., life-threatening) conditions.
Another pillar of ECG automation is the digitization of printed records. While modern ECG machines produce digital signals, patients often receive printed records, limiting the applicability of automated ECG interpretation and ECG-based decision support systems. Digitization, the process of extracting signals from printed records, is challenging due to variations in rotation, paper formats, layouts, and inconsistent lighting in mobile-captured images.
To address these challenges, this work proposes DigiECG: a format-agnostic pipeline that corrects rotation, detects leads, extract signals using a segmentation model, and restores the time scale. The proposed pipeline achieves state-of-the-art performance compared to recent ECG digitization methods, with a Root Mean Squared Error (RMSE) of 52.3 microvolts and a rotation correction RMSE of 0.71 degrees.
